A memory that does not make up who people are
On 24 May, XNeuronal stopped guessing who the people you tell it about are. A person named without their relationship being stated is filed as "social", never as a friend or family by default. Memories said in the same breath are now linked to each other automatically, and the assistant can name a link between two memories: met at, family member of, introduced by.
On Sunday evening, Kate calls her father
Kate calls her father every Sunday, around seven in the evening. He is eighty-three, he lives alone two hours' drive away, and he tells a good story. His week arrives in pieces, in the order it comes back to him: Brenda came round on Tuesday, Frank brought him some bread, the letterbox shuts worse and worse, the plumber is due on Thursday.
What gets lost is what lies between two calls. Who is Brenda? Kate thinks she understood a neighbour, but she is no longer sure. Frank, she had never heard of before last month. From one Sunday to the next the names come back, sometimes in a different order. Sometimes it is her father who gives her a name again as if she did not know it, sometimes it is she who has forgotten it.
So, after hanging up, she presses the button and repeats what she remembers. "Dad saw Brenda on Tuesday. Frank brought him bread. We need to find someone for the letterbox, and the plumber is coming Thursday at nine." She does not say who Brenda and Frank are. She does not really know.
This is exactly the moment when a memory can get things wrong with total confidence.
A missing box: "I don't know yet"
Every memory XNeuronal files is given an area of life. Since 12 May, there had been eight: family, work, health, friends, money, home, learning, self. The area is not decoration. It gives its colour to the dot that stands for the memory on the map, and it serves as a filter when the assistant searches the memory.
For a person, the description the assistant read offered two neighbouring boxes: family, for parents, spouse, children, siblings; and friends, described as "the social circle". No box said "someone whose relationship I don't know". So Brenda had to go somewhere, and the closest thing to "a person who drops by Dad's" was the social circle. Among friends, or in the family, because she moves around a parent.
The flaw did not stop at the area. A memory also carries themes, small keywords that help find it again, and a content, the sentence that sums it up. Nothing prevented writing "friend" in the themes, or summing Brenda up as "your father's friend". Three plausible inventions, and none of them was said. The instructions rewritten that day admit it plainly: it is a serious and frequent mistake. They even give the example of someone tagged "friend" who turns out to be the person who does the cleaning.
It is not just a label, because a memory gets read back. Three weeks later, when Kate asks what she knows about Brenda, the assistant hands back what was filed with the confidence of a memory. A filing error becomes a statement, and nobody checks what they believe they know.
Two more steps: social, then acquaintance
On the morning of 24 May, two areas were added, and they are not like the other eight. They do not say what a memory is about. They say how well you know someone.
"Social" is a person who was only named, whose relationship is unknown. Brenda, this Sunday. "Acquaintance" is someone who comes back, always in the same setting and never outside it: the person you run into at the market every week, the one who drops by on Tuesdays. Above that come friends, a real personal bond outside that setting, then family. Work keeps its own separate place, for colleagues and employers.
The instruction given to the assistant fits in one rule. If the relationship was not stated explicitly, "my wife", "my friend", "my colleague", "my father", it is "social". Never family, friends or work by default. On this scale, it takes the lowest level that what it heard justifies, not the most likely one.
Two rules come with the first. Themes stay empty or strictly factual: "market", "school", but not "friend", "close" or "partner" when nobody stated that relationship. And the content describes what was said: "Brenda, came round to Dad's on Tuesday", or simply "Brenda". The same rule was written into the second pass, the one that rereads each exchange right afterwards to catch the memories a sentence contained without their having been filed.
And when the relationship becomes clearer? The instructions plan a step up. Someone "social" who comes back three or four times in the same setting becomes an "acquaintance". "Actually, she's my sister-in-law" moves someone to "family". This step up goes through the evolution of a memory, the mechanism that never deletes anything: the old memory is archived as history, a new one takes its place, and a "replaced by" link joins the two. We will see at the end that this path had a flaw from that very day.
As long as the relationship is not stated, it stays "social". The second pass sums it up in a sentence you could pin on the wall: it is honest to say "I don't know the relationship yet", it is wrong to pretend to know.
An app that remembers the people I mention on its own, without making up who they are?
That is roughly the question people type into a search engine, and the answer rests on two conditions that only work together. Remembering on its own: Kate opens no contact card and fills in no field. She says "Brenda", and the neural memory creates the person, with what was said about her. Without making things up: what was not said stays empty, and the area says so in plain words.
The difference shows on the day you ask. If Kate asks in November who Brenda is, what was filed allows only one honest answer: someone you told me about, who drops by your father's, whose relationship you have not told me. Not "your father's friend". That blank has value. It tells Kate what she does not know yet, and that may be the question to ask next Sunday.
When you follow the life of a parent who lives alone from a distance, their circle changes without you seeing it, and it is often the family far away who knows least. A memory that mixes up neighbours, friends and family only adds fog. A memory that knows what it does not know at least tells you what to ask.
What is said in the same breath
Back to Kate's sentence. It holds four memories: Brenda's visit on Tuesday, Frank's bread, the letterbox to be repaired, the plumber on Thursday. In their words, they have almost nothing in common: not the same name, not the same place, not the same subject. And yet they belong together. It is her father's week, told on a Sunday evening.
Until then, links between memories were made in two ways. An automatic mechanism compared each new memory with some forty active memories. If they shared enough names, places or themes, with an overlap of at least thirty per cent, or if their meaning was close enough, it made a generic "related to" link, five at most. And the evolution of a memory made its own links, for a correction or a replacement. Neither saw what Kate had just done: say four things with no apparent connection, because they belong to the same story.
Since 24 May, the moment itself makes a link. When a single speaking turn creates at least two memories, whether they come from the assistant during its reply or from the second pass just after, the server links each memory to every other one with a "same turn" link. Four memories, six links. It is not the assistant that decides to make them, it is a server rule, applied at the end of every turn, and the assistant's instructions tell it expressly that it does not have to do this itself.
Two settings say what this link is worth. Its weight is 0.8 out of 1: it is not a kinship of meaning, it is a shared presence. And beyond eight memories created in a single turn, which would be twenty-eight links, the server makes none at all, so as not to weave a web that connects everything to everything and no longer says anything.
This is how you remember things yourself: one detail brings back the whole evening, not the evenings that look like it.
Links that have a name
The automatic link says "it was the same moment". It does not say why two things belong together. For that, on 24 May the assistant received a new tool: linking two existing memories with a typed link, choosing the most precise type that applies.
Six types appeared that day. "Mentioned with", for two people named together. "Met at", for a person and the place where you ran into them. "Talks about", for a conversation or a note and the person or subject it is about. "Family member of" and "colleague of", between two people. "Introduced by", when you got to know someone through a third person. Two older types remain available, "triggered by" for a cause and "concerns" for a broader connection. "Related to" becomes the last resort.
The following Sunday, Kate says: "Frank is Brenda's brother." This time a relationship is stated, and the assistant links Frank to Brenda with "family member of". This is the finest detail of the day: neither Frank nor Brenda changes area. They stay "social" for Kate, because Kate knows them no better. The family bond exists between the two of them, not between them and her. The instructions describe this exact case.
The Sunday after that: "Dad got to know Frank through Brenda." An "introduced by" is added between the same two memories. Nobody said Frank was a friend of her father's, and nothing claims it.
Three safeguards frame the tool. It is only called when a link is known, because the person said it or because it follows explicitly from what they said, never to fill a gap. The two memories must be different. And each link carries a strength between 0 and 1, set to 1 by default for a confirmed link, with 0.6 recommended when it is uncertain.
What Kate sees: the card and the map
All of this would remain invisible plumbing if nothing showed. Two places change.
A memory's card has a "Links" section, with their number. Each line shows an arrow for the direction, the name of the link, then the memory at the other end and its area. Frank's card can therefore show "family member of", Brenda, "social", and below that the memories created in the same sentence, marked "same turn". Kate reads what the app knows about Frank, and she also sees what she does not know.
On the constellation, the two new areas were given deliberately muted colours: a slate grey for "social", the same grey as memories with no area, and a slate one shade darker for "acquaintance". People you know nothing about should not shine like family in amber or friends in green.
The links got their own code too. Warm pink for family, navy blue for work, dotted green for a meeting in a place, wider dotted gold for an introduction, sky blue for "talks about", pale violet for "mentioned with". The "same turn" link is the most discreet of all: a light grey in fine dots, because there are many of them and each one says little.
The next morning, pitch-black dots
On 25 May, first thing, a memory of a person looked odd on the map: a pitch-black disc, with a white circle in the middle and a small dot at the centre. The white circle and the dot are the normal shape of a person on the constellation. The black disc was its halo.
The cause came down to a list. The coloured halos on the map are gradients prepared in advance, one per area, and each dot points to the one for its area. The list of areas for which a gradient was prepared was written by hand in the map component, and it stopped at the eight old ones. Memories filed as "social" or "acquaintance" therefore pointed to a gradient that did not exist, and the drawing library, unable to find the colour, paints it opaque black.
In other words, the previous day's rule was working: the first people filed honestly were precisely the ones that displayed badly.
The fix did more than add two lines. The map now reads the single list of areas kept by the app's memory service, instead of keeping its own copy. No halo will be forgotten again.
What is still open
The most important point concerns the step up. Going from "social" to "acquaintance", or to "family", goes through the evolution of a memory. But evolution, as it was written that day, copies the old memory's area into the new one, and the tool accepts no area as a parameter. The instructions therefore describe a step up the tool could not yet make: the history is chained, the content is updated, the area stays "social".
The second concerns search. Since 12 May, when the assistant finds a memory, it follows its links one hop to bring back neighbouring memories. That hop only follows the generic links, "related to" and "concerns". The named links of 24 May and the "same turn" links show on the card and on the map, and they count when a memory that was found keeps its closest neighbours within reach. But the one-hop search does not take them yet.
The third lies in the nature of the rule. Filing a person as "social" is an instruction given to a model, not a check made by the server. The instruction makes the mistake much rarer, it does not make it impossible.
For Kate, what 24 May changes fits in a few words. On Sunday evening she repeats what her father told her, with the names as they come. What she does not know stays marked as such, what was said together stays together, and on the day someone finally tells her who Frank is, the link is made between the right people. The app knows no more than she does about Brenda. It simply no longer pretends.